[pymvpa] Seg on OSX when using libsvm

[Fabian Pedregosa]

Michael Hanke wrote:
> On Wed, Mar 10, 2010 at 09:40:13AM +0100, Fabian Pedregosa wrote:
>> Yaroslav Halchenko wrote:
>>> Wow -- nice finding Fabian!
>>> I wonder what should be the resolution... and why even the confusion
>>> since shogun's SVC_Q lies in shogun space (not sure how well that is
>>> represented in dynamic library, or may be on your box shogun/libsvm and
>>> pymvpa were built with DYLD_FORCE_FLAT_NAMESPACE or smth like that)
>>>
>>> Unfortunately it isn't the case of the segfaults I can reproduce :-/
>>> (if you have a moment/desire -- try running the unittests extract which
>>> supposed to trigger it)
>> Thanks Yaroslav for your help. It's probably a miscompilation issue
>> on my part (I'm doing some LibSVM development)
> 
> And you gave us some homework ;-)
> 
>   http://fseoane.net/blog/2010/fast-bindings-for-libsvm-in-scikitslearn/
> 
> Thanks for this analysis and the improvements!


It is true that I achieved better performance, but at the long run I'm a 
bit concerned on maintainability, i.e. what will happen when libsvm 
makes a new release. I tried to remain as modular as possible, but as I 
access more fields that those exported in the API, there necessarily 
will be some manual coding in case they decide to modify the internals.

Every binding I've seen has this problem, so in the long run I'm pretty 
sure that either libsvm will have to change their API, either a new svm 
library with a better API will take up its place.


Cheers,

fabian


PS: There's also a very interesting comment pointing to milk [1].

[1] http://luispedro.org/software/milk

> 
> Michael
>